AI model restrictions as a new kind of supply-chain shock
AI model restrictions are government or vendor rules that limit where, how, or by whom advanced models can be deployed, turning intangible software into a strategic supply-chain asset that can be interrupted, rationed, or withdrawn with little warning and wide operational impact. The recent suspension of Anthropic’s Claude Fable 5 and Claude Mythos 5 under a US export-control directive showed how fast this can happen, and how broad the blast radius can be. Anthropic disabled both models globally because screening every user by nationality was not feasible, exposing a new type of risk for AI-native operations. SAP customers using Joule were spared an outage because production workloads did not yet depend on those versions, but the incident made clear that a single directive can remove a core capability, with no migration window, service credits, or advance notice.
Fragmented access: one vendor, different rules by region
Geopolitical AI access is also fragmenting through vendor strategies, especially where model deployment restrictions differ by market. Microsoft’s unique contract with OpenAI allows it to sell GPT models abroad on its own terms, even where OpenAI itself will not sell directly. According to Bloomberg, Azure’s AI revenue in China grew faster than in any other sales territory, roughly tripling in the financial year to June 2025 after climbing about 400% the year before. That growth includes large internet platforms buying GPT through Azure while Anthropic’s models remain absent from the local catalog. The result is an uneven landscape: the same enterprise group may have access to OpenAI models in one region through Microsoft, while facing hard limits or outright bans in another. For global teams, consistency is no longer guaranteed at the model layer.
Regional AI models as a hedge against geopolitical exposure
As AI model restrictions tighten, regional AI models are emerging as a hedge against overdependence on US-controlled systems. OVHcloud’s push to build frontier AI models from scratch is framed by its CEO as a matter of survival, not marketing. The company has already completed pre-training on one model using Europe’s fastest supercomputer and plans a family of models aimed at specific tasks. Octave Klaba argues that advances in chips, training techniques, and synthetic data have reduced the cost of a frontier project from about 1 billion euros to 150–200 million euros. OVHcloud also intends to open-source its models once performance is strong enough, which would give enterprises a transparent, regionally controlled option for sensitive workloads. This kind of regional AI lab positions itself as a substitute when access to American or other foreign models is curtailed.
Enterprise AI supply chain: substitution, continuity and backup plans
For enterprise AI supply chains, the main operational challenge is model substitution when a preferred model becomes unavailable. The Anthropic directive showed that even models not yet in production can be removed without warning, forcing teams to ask what happens if the primary reasoning engine disappears. SAP’s Generative AI Hub offers one answer: governed access to multiple providers, with the option to swap models while keeping a stable integration layer. That pattern—vendor diversification plus a model-agnostic interface—is likely to become standard. Continuity plans now need to include runbooks for switching models, retraining prompts, validating outputs, and handling compliance changes. Procurement and security teams will also need clauses that cover export-control events, not only uptime. AI models can no longer be treated as interchangeable utilities; they are regulated components in a fragile, geopolitical supply chain.
